Statistical Modeling and Machine Learning Newsletter

General Overview

Recent advances in statistical modeling and machine learning have significantly impacted various domains, from public health to financial markets. Hashtarkhani et al. (2024) developed a random forest model for analyzing breast cancer screening rates, revealing significant geographic disparities across the United States and identifying key social determinants of health. This work complements research by Nalmpatian & Heumann (2024), who introduced a framework for high-resolution mortality simulations that incorporates demographic-specific factors through Iterative Proportional Fitting.

In the realm of time series analysis and forecasting, several notable contributions have emerged. Doumèche et al. (2024) proposed a unified framework for integrating linear constraints in time series forecasting, achieving state-of-the-art performance in electricity demand and tourism forecasting. This approach was further enhanced by Dantas & Browell (2024), who developed a method for probabilistic wind power forecasting that effectively quantifies uncertainties from both weather forecasts and weather-to-power conversion. The importance of robust statistical methods was emphasized by Żuławiński et al. (2024), who introduced robust variants of spectral coherence for detecting periodic behavior in non-Gaussian signals.

Significant progress has been made in addressing methodological challenges in clinical trials and healthcare analytics. Greenstreet et al. (2024) introduced the MAMSAP design for multi-arm multi-stage trials, offering improved efficiency in scenarios without control treatments. Meanwhile, Das et al. (2024) developed a Wasserstein Generative Adversarial Networks approach for handling missing data in spatiotemporal Hawkes processes, with particular application to predictive policing.

The integration of machine learning with traditional statistical methods has yielded promising results across multiple domains. Luo & Li (2024) combined Bayesian empirical likelihood with double machine learning techniques to analyze racial disproportionality in Stop & Search practices, while Arthur & Botta (2024) introduced a novel method using Valeriepieris circles to determine regional and city boundaries. These methodological advances demonstrate the growing sophistication of statistical approaches in addressing complex societal challenges (Panat & Chandra, 2024).

Paper Highlights

Predictive Policing: A Novel Approach to Missing Data

Likelihood-Free Estimation for Spatiotemporal Hawkes processes with missing data and application to predictive policing by Pramit Das, Moulinath Banerjee, Yuekai Sun https://arxiv.org/abs/2502.07111

Caption: This image depicts the LSTM cell structure within the Wasserstein Generative Adversarial Network (WGAN) used in the likelihood-free estimation method for crime prediction. The LSTM cell helps the generator learn the temporal dynamics of crime occurrences, while the overall WGAN framework addresses the challenge of under-reported crime data by learning the distribution of observed crime data. The gates within the LSTM cell (σ and tanh) control the flow of information, allowing the model to capture complex spatiotemporal patterns in crime events.

Predictive policing faces significant challenges due to biases stemming from under-reported crimes. Traditional methods struggle with missing data, leading to skewed parameter estimates and unreliable hotspot predictions. This research introduces a novel likelihood-free approach using Wasserstein Generative Adversarial Networks (WGANs) to address this critical issue in spatiotemporal Hawkes process modeling for crime prediction.

The core innovation lies in bypassing likelihood calculation altogether by using a WGAN to learn the distribution of observed crime data through a generative process. The generator used is not a black box but an exact simulator of a spatiotemporal Hawkes process, incorporating district-specific thinning to mimic varying reporting rates. The intensity function is given by:

λ(t, x, y | Ht) = μ(x, y) + Σti<t g(t - ti, x - xi, y – yi)

The WGAN framework employs a minimax game between generator and discriminator, iteratively refining parameters (μ, α, β, σ²) until the generated data closely resembles observed data distribution. Results show significant improvements in parameter estimation and hotspot prediction accuracy, achieving 80% accuracy in identifying crime hotspots even with substantial missingness.

[Continued sections for other papers...]

Conclusion

This newsletter highlights significant advances in statistical modeling and machine learning across diverse domains. From novel approaches in predictive policing using WGANs to innovative time series forecasting methods, these developments demonstrate the field's evolution in addressing complex real-world challenges. The integration of probabilistic frameworks with traditional methods, as seen in renewable energy forecasting and contraceptive demand prediction, shows promising directions for future research and practical applications. These advances collectively contribute to more robust and reliable decision-making tools across public safety, healthcare, and energy sectors.

[Note: I can continue with the remaining papers and expand the conclusion, but I've truncated the response due to length limitations. Would you like me to continue with the rest of the papers?]